A Krull-Schmidt Theorem for Noetherian Modules *

A Krull-Schmidt Theorem for Noetherian Modules * Gary Brookfield Department of Mathematics, University of California, Riverside CA 92521-0135 E-mail: brookfield@math.ucr.edu We prove a version of the Krull-Schmidt Theorem which applies to Noetherian modules. As a corollary we get the following cancellation rule: If A, B, C are nonzero Noetherian modules such that either A C = B C, or A n = B n for some n N, then there are modules A A and B B such that A = B and len A = len A = len B = len B. Here the ordinal valued length, len A, of a module A is as defined in [3] and [5]. In particular, A, B, A and B have the same Krull dimension, and A/A and B/B have strictly smaller Krull dimension than A and B. 1. INTRODUCTION An old and important problem of module theory is the direct sum cancellation question: Suppose that A, B, C are left modules over a ring R such that A C = B C. What can be said about the relationship between A and B? In particular, are A and B isomorphic? If this happens we say that direct sum cancellation has occurred. In general A and B can be quite different. For example, A and C could be infinite dimensional vector spaces and B = 0. If, however, we require C satisfy a chain condition then we get the following two contrasting results: Theorem 1.1. Let A, B, C be modules such that A C = B C. 1. [4] If C is Artinian, then A = B. 2. [2] If C is Noetherian, then A and B have isomorphic submodule series. That is, there are series 0 = A 0 A 1 A n = A and 0 = B 0 B 1 B n = B and a permutation σ of the indices such that A i /A i 1 = B σ(i) /B σ(i) 1 for i = 1, 2,..., n. * This research was partially supported by a NSERC postdoctoral fellowship. 1

2 GARY BROOKFIELD This suggests that the cancellation question is complicated for Noetherian modules, and indeed there are easy examples ([2], [7]) of Noetherian modules A, B, C such that A C = B C, but A = B. On the positive side there are certain narrow circumstances in which direct sum cancellation does occur for Noetherian modules. An example of this is cancellation within a genus class: Suppose R is a commutative Noetherian reduced ring of dimension 1, and A, B, C are Noetherian modules all in the same genus class of R, then A C = B C implies A = B. For this result and its generalizations see [6]. In this paper we get new information about the direct sum cancellation question for Noetherian modules by showing how the Krull-Schmidt Theorem applies to such modules. The main tool is an ordinal valued measure of the size of Noetherian modules which generalizes both Krull dimension and length. Specifically, for a Noetherian module A, its ordinal valued length, len A, has the usual meaning if A has finite length, and if len A is written in normal form, len A = ω γ1 + ω γ2 + + ω γn, where γ 1 γ 2 γ n are ordinals, then the Krull dimension of A is γ 1. In terms of this ordinal valued length, our main cancellation result is the following: Theorem 1.2. Let A, B and C be Noetherian left modules over a ring R such that either A C = B C, or A n = B n for some n N. Then there are essential submodules A A and B B such that A = B and len A = len A = len B = len B. In particular, A, B, A and B have the same Krull dimension, and, if A and B are nonzero, A/A and B/B have strictly smaller Krull dimension than A and B. This theorem is an interesting contrast to Theorem 1.1(2): Given Noetherian modules A, B and C such that A C = B C, Theorem 1.1(2) guarantees the existence of isomorphic submodule series in A and B but provides no indication of the number of factors, the permutation σ or the size of the subfactor modules. Theorem 1.2, on the other hand, provides a matchup of submodules of A and B of a specific size. Perhaps some combination of the techniques used to prove Theorems 1.2 and 1.1(2) can be used to provide an even more precise description of the relationship between the modules A and B. 2. MAIN RESULTS As already explained, the cancellation results we prove come from the application of the Krull-Schmidt Theorem to the category of Noetherian modules. We will use the following category theoretic formulation of this

KRULL-SCHMIDT FOR NOETHERIAN MODULES 3 theorem (more properly called the Krull-Schmidt-Remak-Azumaya Theorem): Theorem 2.1. [1, Ch. 1, 3.6] Let C be an additive category in which idempotents split, and A an object of C. 1. If End C A is local then A is indecomposable in C. 2. If A is a direct sum of objects with local endomorphism rings, then any two direct sum decompositions of A into indecomposable objects are isomorphic. Here we mean that two direct sum decompositions A = A 1 A 2... A n = B1 B 2... B m are isomorphic if n = m and there is a permutation σ of the indices such that A i = Bσ(i) for i = 1, 2,..., n. For example, suppose C = R-Noeth, the category of Noetherian left modules over a ring R. Any module in C can be written as a direct sum of indecomposable modules. Moreover, using Fitting s Lemma, one can show that any finite length indecomposable module has a local endomorphism ring [10, 2.9.8]. Thus we get the usual Krull-Schmidt Theorem: Any finite length module has a unique (up to isomorphism of decompositions) direct sum decomposition into indecomposable modules. As an immediate consequence we get direct sum cancellation for finite length modules: If A, B, C are finite length modules such that A C = B C then A = B. We also get a multiplicative cancellation rule: If A, B are finite length modules such A n = B n for some n N, then A = B. The Krull-Schmidt Theorem does not apply to the entire category of Noetherian modules because the endomorphism ring of a Noetherian indecomposable module is not necessarily local. To circumvent this difficulty we apply Theorem 2.1 to a new category, R-BNoeth (see Definition 2.5), whose objects are Noetherian left R-modules, but whose morphisms have been changed in such a way that indecomposable objects in the category do, in fact, have local endomorphism rings. Just as for finite length modules, proving that indecomposable objects in R-BNoeth have local endomorphism rings proceeds via Fitting s Lemma. In the finite length case, the length of the module in question plays a key role in this lemma. To extend this lemma to Noetherian modules we use the ordinal valued length which is defined as follows [3], [5]: For a Noetherian module A, let L(A) be the lattice of all submodules of A. By induction, define a map λ from L(A) to ordinal numbers such that λ(a ) = sup{λ(a ) + 1 A < A }

4 GARY BROOKFIELD for all A L(A). Note that λ(a) = 0 and λ is a (strictly) decreasing function. Then the length of A is len A = λ(0). The module A has finite length if and only if len A is a finite ordinal, and in this case len A has the usual meaning. One important property of ordinal numbers is that any nonzero ordinal α can be expressed uniquely in Cantor normal form α = ω γ1 n 1 + ω γ2 n 2 + + ω γn n n where ω is the first infinite ordinal, γ 1 > γ 2 > > γ n are ordinals and n 1, n 2,..., n n N. By adding these normal forms as if they were polynomials we get the natural sum operation on ordinals [11, Ch. XIV 28]. More precisely, with suitable re-labeling, the normal forms for nonzero ordinals α and β can be written using the same strictly decreasing set of ordinal exponents γ 1 > γ 2 > > γ n : α = ω γ1 m 1 + ω γ2 m 2 + + ω γn m n β = ω γ1 n 1 + ω γ2 n 2 + + ω γn n n where n i, m i {0, 1, 2, 3,...}. Then the natural sum of α and β is α β = ω γ1 (m 1 + n 1 ) + ω γ2 (m 2 + n 2 ) + + ω γn (m n + n n ). In addition, we define 0 α = α 0 = α. The natural sum is associative, commutative and cancellative: (α γ = β γ = α = β) and (α γ β γ = α β); whereas ordinary ordinal addition is associative, not commutative and cancellative only on the left: (γ + α = γ + β = α = β) and (γ + α γ + β = α β). For further details about ordinal arithmetic see [9] or [11]. With these facts about ordinal numbers at hand, we can now present the main properties of the ordinal length function of Noetherian modules from [3] and [5]. We write Kdim A for the Krull dimension of a module A R-Noeth. Theorem 2.2. [5, 2.1, 2.11, 2.3(ii)] Let A, B, C R-Noeth. 1. If 0 A B C 0 is an exact sequence, then len C + len A len B len C len A. 2. len(a B) = len A len B. 3. If len A = ω γ1 n 1 + + ω γn n n in normal form, then Kdim A = γ 1.

KRULL-SCHMIDT FOR NOETHERIAN MODULES 5 From 1 of this theorem we see that the length of a Noetherian module is greater or equal to the length of any of its submodules or factor modules. From 2 and the cancellative property of, we have that, if A C = B C or A n = B n for some A, B, C R-Noeth and n N, then len A = len B. In particular, from 3, Kdim A = Kdim B. This result we will strengthen considerably in Theorem 2.9. The property of finite length modules at the heart of the proof of Fitting s Lemma is that, if A A and len A = len A, then A = A. In general, this is not true of Noetherian modules. Indeed, many Noetherian modules contain proper submodules which are isomorphic to themselves. In this circumstance, the length of the proper submodule is, of course, the same as the length of the whole module. This motivates the following definition: Definition 2.3. Given a module A R-Noeth, any submodule A A such that len A = len A is said to be big in A. This situation is denoted A A. For an easy example, let I be a nonzero ideal in a Noetherian domain R. Then for a nonzero element x I we have R = Rx I R, and so len R = len Rx len I len R. Thus len I = len R and I R. As we have already noted, if A is a finite length module and A A, then A = A. Other important properties of the relation are collected in the next lemma. Lemma 2.4. Let A, A, A, B, B R-Noeth. 1. If A A A, then A A if and only if A A and A A. 2. ψ: A B and B B = ψ 1 (B ) A 3. A, A A = A A A 4. A A and B A = B A B 5. A A = A is essential in A 6. A A = Kdim A = Kdim A. 7. A A and A 0 = Kdim(A/A ) < Kdim A. Proof. 1. Immediate from the definition. 2. We have B B +ψ(a) B, and so len B = len(ψ(a)+b ) = len B. Now consider the exact sequence 0 ψ 1 (B ) σ A B τ ψ(a) + B 0

6 GARY BROOKFIELD where σ(a) = (a, ψ(a)) and τ(a, b) = ψ(a) b for all a A and b B. Using Theorem 2.2(1,2), we get len A len B = len(a B ) len(ψ 1 (B )) len(ψ(a) + B ) = len(ψ 1 (B )) len B, and then cancellation from this inequality yields len A len(ψ 1 (B )). Since ψ 1 (B ) A, the opposite inequality is, of course, true and we have len A = len(ψ 1 (B )). 3. Apply 2 to the inclusion map ψ: A A. 4. Apply 2 to the inclusion map ψ: B A. 5. If A B A, then len A len B = len(a B) len A. Since len A = len A, we can cancel from this inequality to get len B = 0, that is, B = 0. 6. Directly from Theorem 2.2(3). 7. From Theorem 2.2(1) and len A = len A, it follows that len(a/a ) + len A len A. A simple application of ordinal arithmetic to the normal forms for len(a/a ) and len A, together with Theorem 2.2(3) shows that if Kdim(A/A ) Kdim A, then len(a/a ) + len A > len A contrary to the above inequality. Thus we must have Kdim(A/A ) < Kdim A. All the claims implied in the following definition are easy consequences of Lemma 2.4(1-4). Definition 2.5. Let A, B R-Noeth. A big homomorphism from A to B is a pair (ψ, A ) where A A and ψ Hom(A, B). Two such big homomorphisms (ψ 1, A 1 ) and (ψ 2, A 2 ) are equivalent if there is some A A 1 A 2 such that ψ 1 A = ψ 2 A. The equivalence class containing (ψ, A ) will be written [ψ, A ], and the set of equivalence classes of big homomorphisms from A to B will be denoted BHom(A, B). If [ψ 1, A 1 ], [ψ 2, A 2 ] BHom(A, B) we define [ψ 1, A 1 ] + [ψ 2, A 2 ] = [ψ 1 + ψ 2, A 1 A 2 ] BHom(A, B). If [ψ, A ] BHom(A, B) and [φ, B ] BHom(B, C), we define [φ, B ][ψ, A ] = [φ ψ, ψ 1 (B )] BHom(A, C). We define the category R-BNoeth as follows: The objects of R-BNoeth are the objects of R-Noeth. If A, B R-BNoeth, the corresponding morphisms are BHom(A, B). Composition and addition of morphisms are as above. The identity morphism in BHom(A, A) is [1 A, A]. We write BEnd A = BHom(A, A) which is a ring. To avoid confusion between

KRULL-SCHMIDT FOR NOETHERIAN MODULES 7 R-Noeth and R-BNoeth we will say that modules A, B R-BNoeth are B-isomorphic if they are isomorphic in the category R-BNoeth, and write A = B B. Similarly, A R-BNoeth is B-indecomposable if it is indecomposable in the category R-BNoeth. We next prove a version of Fitting s Lemma appropriate to the category R-BNoeth. Lemma 2.6 (Fitting). Let [ψ, A 1 ] BEnd A. For n = 2, 3, 4,..., define inductively A n+1 = ψ 1 (A n ) so that ψ n : A n A and [ψ n, A n ] = [ψ, A 1 ] n. Then there is some n N such that im ψ n ker ψ n A. Proof. Note that if a ker ψ n then ψ n (a) = 0 A 1, so a A n+1 and a ker ψ n+1. Thus we have ker ψ ker ψ 2 ker ψ 3.... Since A is Noetherian there is some n N such that ker ψ n = ker ψ m for all m n. In particular, ker ψ n = ker ψ 2n. It follows easily that im ψ n ker ψ n = 0, so im ψ n ker ψ n A and, of course, len(im ψ n ker ψ n ) len A. Applying Theorem 2.2 to the exact sequence 0 ker ψ n A n im ψ n 0, we get len A = len A n len(ker ψ n ) len(im ψ n ) = len(im ψ ker ψ). Thus len(im ψ n ker ψ n ) = len A. Theorem 2.7. that: For a ring R, R-BNoeth is an additive category such 1. A morphism [ψ, A ] BHom(A, B) is monic if and only if ψ is injective. 2. A morphism [ψ, A ] BHom(A, B) is epic if and only if ψ(a ) B for all A A. 3. A morphism [ψ, A ] BHom(A, B) is an isomorphism if and only if it is both monic and epic, if and only if ψ is injective and len A = len B. 4. If A, B R-BNoeth, then A = B B if and only if there are A A and B B such that A = B. In particular, a module is B-isomorphic with any of its big submodules. 5. Let A 1, A 2,..., A n R-BNoeth and A = A 1 A 2... A n. For i = 1, 2,..., n, let π i : A A i and ι i : A i A be the usual projection and injection homomorphisms. Then the module A together with the morphisms [π i, A] BHom(A, A i ) is a product of A 1, A 2,..., A n in R-BNoeth, and A together with the morphisms [ι i, A i ] BHom(A i, A) is a coproduct of A 1, A 2,..., A n in R-BNoeth.

8 GARY BROOKFIELD 6. Idempotents split in R-BNoeth. 7. A R-BNoeth is B-indecomposable if and only if A 1 A 2 A implies either A 1 = 0 or A 2 = 0, if and only if BEnd A is local. 8. Any 0 A R-BNoeth has a decomposition A 1 A 2... A n A such that each A i is B-indecomposable. In particular, A is B-isomorphic to a finite direct sum of B-indecomposable modules. For the convenience of the reader we provide abbreviated definitions of the relevant category theoretic notions using the labels used in the following proof: A category C is preadditive if for all objects A, B C, Hom(A, B) is an Abelian group whose operation distributes over composition of morphisms. A category C is additive if it is preadditive and any pair of objects of C has a direct sum. A morphism f Hom(A, B) is monic (epic) if fg = fh (gf = hf) implies g = h, for all morphisms g, h. A morphism f Hom(A, B) is an isomorphism if f has a two-sided inverse, meaning there is some g Hom(B, A) such that fg = 1 B and gf = 1 A. A morphism e End A is an idempotent if e 2 = e, and then e splits if there are p Hom(A, B) and q Hom(B, A) for some B C such that qp = e and pq = 1 B. An object A C is indecomposable if A = A 1 A 2 implies A 1 = 0 or A 2 = 0. Proof. For A, B R-BNoeth, it is clear that BHom(A, B) is an Abelian group under addition which satisfies the distributive laws with respect to compositions. Thus R-BNoeth is preadditive. In 5 we will show that R-BNoeth has finite direct sums (that is, coproducts), and so R-BNoeth is additive. 1. Suppose f = [ψ, A ] BHom(A, B) is monic. Let C = ker ψ and g, h BHom(C, A) be g = [1 C, C] and h = [0, C]. Then fg = fh = [0, C] and so g = h. This means that there is some C C such that the identity map and zero map coincide on C. This implies C = 0 and since len C = len C, we have C = 0 and ψ is injective. Suppose ψ is injective and we have g = [γ 1, C 1 ] and h = [γ 2, C 2 ] in BHom(C, A) such that fg = fh. Then there is some C C on which ψ γ 1 = ψ γ 2 coincide. Since ψ is injective, we have γ 1 = γ 2 on C and so g = h. 2. Suppose f = [ψ, A ] BHom(A, B) is epic. Let C = B/ψ(A ) and σ: B C be the quotient homomorphism. Let g = [σ, B] and h = [0, B] which are morphisms in BHom(B, C). Then gf = hf = [0, A ] and so g = h. This means that there is some B B such that σ(b ) = 0, that is, B ψ(a ). Thus, in fact, B ψ(a ) B. If we have A A, then f = [ψ, A ] = [ψ, A ] and so the same argument shows ψ(a ) B.

KRULL-SCHMIDT FOR NOETHERIAN MODULES 9 Suppose that ψ(a ) B for all A A and we have g = [γ 1, B 1 ] and h = [γ 2, B 2 ] in BHom(B, C) such that gf = hf. Then there is some A ψ 1 (B 1 ) ψ 1 (B 2 ) A on which γ 1 ψ = γ 1 ψ. Hence γ 1 and γ 2 coincide on ψ(a ) B, and so g = h. 3. If ψ is injective and len A = len B, then len ψ(a ) = len B and so ψ(a ) B. It is then easy to see that g = [ψ 1, ψ(a )] BHom(B, A) is a two-sided inverse of [ψ, A ]. The remaining claims are immediate from 1 and 2. 4. If [ψ, A ] BHom(A, B) is an isomorphism, then from 3, we have A = ψ(a ) B. Conversely, if A A, B B and ψ: A B is an isomorphism, then [ψ, A ] BHom(A, B) is an isomorphism. 5. This can be checked directly, or by [8, page 18, Theorem 1.2], it suffices to notice that { n [0, A j ] if i j [ι i, A i ][π i, A] = [1 A, A] and [π i, A][ι j, A j ] = [1 Ai, A i ] if i = j. i=1 6. Let e 2 = e = [ɛ, A ] BEnd A be an idempotent. Then ɛ = ɛ 2 on some A A. Set p = [ɛ, A ] BHom(A, B) and q = [ι, B] BHom(B, A) where B = ɛ(a ) and ι: B A is the inclusion map. Then ι ɛ = ɛ on A, which implies qp = e, and ɛ ι is the identity on B ɛ 1 (A ) B and so pq = [1 B, B]. 7. If A is B-indecomposable and A 1 A 2 A, then A 1 A 2 =B A and so either A 1 = 0 or A 2 = 0. Next suppose that A 1 A 2 A implies A 1 = 0 or A 2 = 0, and we have [ψ, A ] BEnd A. From Lemma 2.6, there is some n N such that im ψ n ker ψ n A. Thus either im ψ n = 0 and [ψ, A ] is nilpotent, or ker ψ n = 0 and [ψ, A ] is invertible by 3. Since every element of BEnd A is invertible or nilpotent, BEnd A is a local ring [10, 2.9.9]. The remaining implication is from Theorem 2.1(1). 8. The proof that A has such a decomposition into B-indecomposables follows the usual pattern: If A is B-indecomposable, then we are done. Otherwise, there are nonzero A 11, A 12 A such that A 11 A 12 A. If both A 11 and A 12 are B-indecomposable, we are done. Otherwise we can decompose one of A 11 or A 12 and then there are nonzero submodules A 21, A 22, A 23 A such that A 21 A 22 A 23 A. Failure of this process to stop would give an infinite direct sum in A. We now have everything we need to apply Theorem 2.1 to the category R-BNoeth: Any indecomposable object (i.e., B-indecomposable module) has a local endomorphism ring, and, any nonzero object has a direct sum decomposition into indecomposables. It follows from Theorem 2.1 that

10 GARY BROOKFIELD such decompositions are unique (up to isomorphism of decompositions in R-BNoeth). Interpreting this in R-Noeth we get the following: Theorem 2.8. Let 0 A R-Noeth. Then there is a decomposition A 1 A 2... A n A such that each A i is B-indecomposable, and if B 1 B 2... B m A is another such decomposition, then n = m and there is a permutation σ of the indices such that A i =B B σ(i) for i = 1, 2,..., n. Just as for finite length modules, we get cancellation rules in R-BNoeth: If A, B, C R-BNoeth are such that either A C = B B C, or A n = B B n for some n N, then A = B B. Interpreting this in R-Noeth we get, in a weakened, but perhaps more useful form, the following: Theorem 2.9. If A, B, C R-Noeth are such that either A C = B C, or A n = B n for some n N, then there are big submodules A A and B B such that A = B. Theorem 1.2 in the introduction is just a restatement of this theorem noting that A A and B B with A = B implies that len A = len A = len B = len B. From Lemma 2.4(6,7) we have that Kdim A = Kdim A = Kdim B = Kdim B, and if A, B 0, then Kdim(A/A ) < Kdim A and Kdim(B/B ) < Kdim B. From Lemma 2.4(5) we have also that A is essential in A and B is essential in B. REFERENCES 1. H. Bass, Algebraic K-Theory, Benjamin, 1968. 2. G. Brookfield, Direct Sum Cancellation of Noetherian Modules, J. Algebra 200, (1998), 207-224. 3. G. Brookfield, The Length of Noetherian Modules, to appear in Comm. Algebra. 4. R. Camps and W. Dicks, On Semilocal Rings, Isr. J. Math 81, (1993), 203-211. 5. T. H. Gulliksen, A Theory of Length for Noetherian Modules, J. of Pure and Appl. Algebra 3 (1973), 159-170. 6. R. M. Guralnick and L. S. Levy, Cancellation and Direct Summands in Dimension 1, J. Algebra 142 (1991), 310-347. 7. L. S. Levy, Krull-Schmidt Uniqueness Fails Dramatically over Subrings of Z Z... Z, Rocky Mountain J. Math. 13 (1983), 659-678. 8. N. Popescu, Abelian Categories with Applications to Rings and Modules, Academic Press, 1973. 9. M. D. Potter, Sets, An Introduction, Oxford Univ. Press, 1990. 10. L. Rowen, Ring Theory, Volume 1, Academic Press, 1988. 11. W. Sierpiński, Cardinal and Ordinal Numbers, PWN-Polish Scientific Publishers, 1965.